zephyr-7b-dpo-full-ultrabin-2

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5012
• Rewards/chosen: -1.0205
• Rewards/rejected: -1.9949
• Rewards/accuracies: 0.7734
• Rewards/margins: 0.9744
• Logps/rejected: -462.1516
• Logps/chosen: -364.6655
• Logits/rejected: 0.2960
• Logits/chosen: -0.5539

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5737	0.2092	100	0.5691	-0.6944	-1.2209	0.7227	0.5264	-384.7459	-332.0635	-1.3889	-1.5290
0.5456	0.4184	200	0.5217	-0.8959	-1.6744	0.7812	0.7785	-430.1032	-352.2071	-0.6808	-1.2567
0.4938	0.6276	300	0.5073	-0.9394	-1.8266	0.7695	0.8872	-445.3213	-356.5596	-0.2199	-0.9537
0.5014	0.8368	400	0.5015	-1.0166	-1.9608	0.7656	0.9442	-458.7432	-364.2830	0.1887	-0.6221

Framework versions

• Transformers 4.41.2
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1